lustre/ldlm/ldlm_resource.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2010, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/ldlm/ldlm_resource.c
  33  *
  34  * Author: Phil Schwan <phil@clusterfs.com>
  35  * Author: Peter Braam <braam@clusterfs.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_LDLM
  39 #include <lustre_dlm.h>
  40 #include <lustre_fid.h>
  41 #include <obd_class.h>
  42 #include "ldlm_internal.h"
  43
  44 struct kmem_cache *ldlm_resource_slab, *ldlm_lock_slab;
  45 struct kmem_cache *ldlm_interval_tree_slab;
  46 struct kmem_cache *ldlm_inodebits_slab;
  47
  48 int ldlm_srv_namespace_nr = 0;
  49 int ldlm_cli_namespace_nr = 0;
  50
  51 DEFINE_MUTEX(ldlm_srv_namespace_lock);
  52 LIST_HEAD(ldlm_srv_namespace_list);
  53
  54 DEFINE_MUTEX(ldlm_cli_namespace_lock);
  55 /* Client Namespaces that have active resources in them.
  56  * Once all resources go away, ldlm_poold moves such namespaces to the
  57  * inactive list */
  58 LIST_HEAD(ldlm_cli_active_namespace_list);
  59 /* Client namespaces that don't have any locks in them */
  60 LIST_HEAD(ldlm_cli_inactive_namespace_list);
  61
  62 static struct dentry *ldlm_debugfs_dir;
  63 static struct dentry *ldlm_ns_debugfs_dir;
  64 struct dentry *ldlm_svc_debugfs_dir;
  65
  66 /* during debug dump certain amount of granted locks for one resource to avoid
  67  * DDOS. */
  68 static unsigned int ldlm_dump_granted_max = 256;
  69
  70 static ssize_t ldebugfs_dump_ns_seq_write(struct file *file,
  71                                           const char __user *buffer,
  72                                           size_t count, loff_t *off)
  73 {
  74         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
  75         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
  76         RETURN(count);
  77 }
  78
  79 LDEBUGFS_FOPS_WR_ONLY(ldlm, dump_ns);
  80
  81 static int ldlm_rw_uint_seq_show(struct seq_file *m, void *v)
  82 {
  83         seq_printf(m, "%u\n", *(unsigned int *)m->private);
  84         return 0;
  85 }
  86
  87 static ssize_t
  88 ldlm_rw_uint_seq_write(struct file *file, const char __user *buffer,
  89                        size_t count, loff_t *off)
  90 {
  91         struct seq_file *seq = file->private_data;
  92
  93         if (!count)
  94                 return 0;
  95
  96         return kstrtouint_from_user(buffer, count, 0,
  97                                     (unsigned int *)seq->private);
  98 }
  99
 100 LDEBUGFS_SEQ_FOPS(ldlm_rw_uint);
 101
 102 #ifdef HAVE_SERVER_SUPPORT
 103
 104 static int seq_watermark_show(struct seq_file *m, void *data)
 105 {
 106         seq_printf(m, "%llu\n", *(__u64 *)m->private);
 107         return 0;
 108 }
 109
 110 static ssize_t seq_watermark_write(struct file *file,
 111                                    const char __user *buffer, size_t count,
 112                                    loff_t *off)
 113 {
 114         struct seq_file *m = file->private_data;
 115         u64 value;
 116         __u64 watermark;
 117         __u64 *data = m->private;
 118         bool wm_low = (data == &ldlm_reclaim_threshold_mb) ? true : false;
 119         char kernbuf[22] = "";
 120         int rc;
 121
 122         if (count >= sizeof(kernbuf))
 123                 return -EINVAL;
 124
 125         if (copy_from_user(kernbuf, buffer, count))
 126                 return -EFAULT;
 127         kernbuf[count] = 0;
 128
 129         rc = sysfs_memparse(kernbuf, count, &value, "MiB");
 130         if (rc < 0) {
 131                 CERROR("Failed to set %s, rc = %d.\n",
 132                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb",
 133                        rc);
 134                 return rc;
 135         } else if (value != 0 && value < (1 << 20)) {
 136                 CERROR("%s should be greater than 1MB.\n",
 137                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb");
 138                 return -EINVAL;
 139         }
 140         watermark = value >> 20;
 141
 142         if (wm_low) {
 143                 if (ldlm_lock_limit_mb != 0 && watermark > ldlm_lock_limit_mb) {
 144                         CERROR("lock_reclaim_threshold_mb must be smaller than "
 145                                "lock_limit_mb.\n");
 146                         return -EINVAL;
 147                 }
 148
 149                 *data = watermark;
 150                 if (watermark != 0) {
 151                         watermark <<= 20;
 152                         do_div(watermark, sizeof(struct ldlm_lock));
 153                 }
 154                 ldlm_reclaim_threshold = watermark;
 155         } else {
 156                 if (ldlm_reclaim_threshold_mb != 0 &&
 157                     watermark < ldlm_reclaim_threshold_mb) {
 158                         CERROR("lock_limit_mb must be greater than "
 159                                "lock_reclaim_threshold_mb.\n");
 160                         return -EINVAL;
 161                 }
 162
 163                 *data = watermark;
 164                 if (watermark != 0) {
 165                         watermark <<= 20;
 166                         do_div(watermark, sizeof(struct ldlm_lock));
 167                 }
 168                 ldlm_lock_limit = watermark;
 169         }
 170
 171         return count;
 172 }
 173
 174 static int seq_watermark_open(struct inode *inode, struct file *file)
 175 {
 176         return single_open(file, seq_watermark_show, inode->i_private);
 177 }
 178
 179 static const struct file_operations ldlm_watermark_fops = {
 180         .owner          = THIS_MODULE,
 181         .open           = seq_watermark_open,
 182         .read           = seq_read,
 183         .write          = seq_watermark_write,
 184         .llseek         = seq_lseek,
 185         .release        = lprocfs_single_release,
 186 };
 187
 188 static int seq_granted_show(struct seq_file *m, void *data)
 189 {
 190         seq_printf(m, "%llu\n", percpu_counter_sum_positive(
 191                    (struct percpu_counter *)m->private));
 192         return 0;
 193 }
 194
 195 static int seq_granted_open(struct inode *inode, struct file *file)
 196 {
 197         return single_open(file, seq_granted_show, inode->i_private);
 198 }
 199
 200 static const struct file_operations ldlm_granted_fops = {
 201         .owner  = THIS_MODULE,
 202         .open   = seq_granted_open,
 203         .read   = seq_read,
 204         .llseek = seq_lseek,
 205         .release = seq_release,
 206 };
 207
 208 #endif /* HAVE_SERVER_SUPPORT */
 209
 210 static struct lprocfs_vars ldlm_debugfs_list[] = {
 211         { .name =       "dump_namespaces",
 212           .fops =       &ldlm_dump_ns_fops,
 213           .proc_mode =  0222 },
 214         { .name =       "dump_granted_max",
 215           .fops =       &ldlm_rw_uint_fops,
 216           .data =       &ldlm_dump_granted_max },
 217 #ifdef HAVE_SERVER_SUPPORT
 218         { .name =       "lock_reclaim_threshold_mb",
 219           .fops =       &ldlm_watermark_fops,
 220           .data =       &ldlm_reclaim_threshold_mb },
 221         { .name =       "lock_limit_mb",
 222           .fops =       &ldlm_watermark_fops,
 223           .data =       &ldlm_lock_limit_mb },
 224         { .name =       "lock_granted_count",
 225           .fops =       &ldlm_granted_fops,
 226           .data =       &ldlm_granted_total },
 227 #endif
 228         { NULL }
 229 };
 230
 231 int ldlm_debugfs_setup(void)
 232 {
 233         int rc;
 234
 235         ENTRY;
 236         ldlm_debugfs_dir = debugfs_create_dir(OBD_LDLM_DEVICENAME,
 237                                              debugfs_lustre_root);
 238         ldlm_ns_debugfs_dir = debugfs_create_dir("namespaces",
 239                                                  ldlm_debugfs_dir);
 240         ldlm_svc_debugfs_dir = debugfs_create_dir("services",
 241                                                   ldlm_debugfs_dir);
 242
 243         rc = ldebugfs_add_vars(ldlm_debugfs_dir, ldlm_debugfs_list, NULL);
 244         if (rc != 0) {
 245                 CERROR("LProcFS failed in ldlm-init\n");
 246                 GOTO(err, rc);
 247         }
 248
 249         RETURN(0);
 250
 251 err:
 252         debugfs_remove_recursive(ldlm_debugfs_dir);
 253         ldlm_svc_debugfs_dir = NULL;
 254         ldlm_ns_debugfs_dir = NULL;
 255         ldlm_debugfs_dir = NULL;
 256         RETURN(rc);
 257 }
 258
 259 void ldlm_debugfs_cleanup(void)
 260 {
 261         debugfs_remove_recursive(ldlm_debugfs_dir);
 262
 263         ldlm_svc_debugfs_dir = NULL;
 264         ldlm_ns_debugfs_dir = NULL;
 265         ldlm_debugfs_dir = NULL;
 266 }
 267
 268 static ssize_t resource_count_show(struct kobject *kobj, struct attribute *attr,
 269                                    char *buf)
 270 {
 271         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 272                                                  ns_kobj);
 273         __u64                   res = 0;
 274         int                     i;
 275
 276         /* result is not strictly consistant */
 277         for (i = 0; i < (1 << ns->ns_bucket_bits); i++)
 278                 res += atomic_read(&ns->ns_rs_buckets[i].nsb_count);
 279         return sprintf(buf, "%lld\n", res);
 280 }
 281 LUSTRE_RO_ATTR(resource_count);
 282
 283 static ssize_t lock_count_show(struct kobject *kobj, struct attribute *attr,
 284                                char *buf)
 285 {
 286         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 287                                                  ns_kobj);
 288         __u64                   locks;
 289
 290         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
 291                                         LPROCFS_FIELDS_FLAGS_SUM);
 292         return sprintf(buf, "%lld\n", locks);
 293 }
 294 LUSTRE_RO_ATTR(lock_count);
 295
 296 static ssize_t lock_unused_count_show(struct kobject *kobj,
 297                                       struct attribute *attr,
 298                                       char *buf)
 299 {
 300         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 301                                                  ns_kobj);
 302
 303         return sprintf(buf, "%d\n", ns->ns_nr_unused);
 304 }
 305 LUSTRE_RO_ATTR(lock_unused_count);
 306
 307 static ssize_t lru_size_show(struct kobject *kobj, struct attribute *attr,
 308                              char *buf)
 309 {
 310         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 311                                                  ns_kobj);
 312         __u32 *nr = &ns->ns_max_unused;
 313
 314         if (ns_connect_lru_resize(ns))
 315                 nr = &ns->ns_nr_unused;
 316         return sprintf(buf, "%u\n", *nr);
 317 }
 318
 319 static ssize_t lru_size_store(struct kobject *kobj, struct attribute *attr,
 320                               const char *buffer, size_t count)
 321 {
 322         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 323                                                  ns_kobj);
 324         unsigned long tmp;
 325         int lru_resize;
 326         int err;
 327
 328         if (strncmp(buffer, "clear", 5) == 0) {
 329                 CDEBUG(D_DLMTRACE,
 330                        "dropping all unused locks from namespace %s\n",
 331                        ldlm_ns_name(ns));
 332                 if (ns_connect_lru_resize(ns)) {
 333                         /* Try to cancel all @ns_nr_unused locks. */
 334                         ldlm_cancel_lru(ns, ns->ns_nr_unused, 0,
 335                                         LDLM_LRU_FLAG_PASSED |
 336                                         LDLM_LRU_FLAG_CLEANUP);
 337                 } else {
 338                         tmp = ns->ns_max_unused;
 339                         ns->ns_max_unused = 0;
 340                         ldlm_cancel_lru(ns, 0, 0, LDLM_LRU_FLAG_PASSED |
 341                                         LDLM_LRU_FLAG_CLEANUP);
 342                         ns->ns_max_unused = tmp;
 343                 }
 344                 return count;
 345         }
 346
 347         err = kstrtoul(buffer, 10, &tmp);
 348         if (err != 0) {
 349                 CERROR("lru_size: invalid value written\n");
 350                 return -EINVAL;
 351         }
 352         lru_resize = (tmp == 0);
 353
 354         if (ns_connect_lru_resize(ns)) {
 355                 if (!lru_resize)
 356                         ns->ns_max_unused = (unsigned int)tmp;
 357
 358                 if (tmp > ns->ns_nr_unused)
 359                         tmp = ns->ns_nr_unused;
 360                 tmp = ns->ns_nr_unused - tmp;
 361
 362                 CDEBUG(D_DLMTRACE,
 363                        "changing namespace %s unused locks from %u to %u\n",
 364                        ldlm_ns_name(ns), ns->ns_nr_unused,
 365                        (unsigned int)tmp);
 366                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 367
 368                 if (!lru_resize) {
 369                         CDEBUG(D_DLMTRACE,
 370                                "disable lru_resize for namespace %s\n",
 371                                ldlm_ns_name(ns));
 372                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
 373                 }
 374         } else {
 375                 CDEBUG(D_DLMTRACE,
 376                        "changing namespace %s max_unused from %u to %u\n",
 377                        ldlm_ns_name(ns), ns->ns_max_unused,
 378                        (unsigned int)tmp);
 379                 ns->ns_max_unused = (unsigned int)tmp;
 380                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_LRU_FLAG_PASSED);
 381
 382                 /* Make sure that LRU resize was originally supported before
 383                  * turning it on here.
 384                  */
 385                 if (lru_resize &&
 386                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
 387                         CDEBUG(D_DLMTRACE,
 388                                "enable lru_resize for namespace %s\n",
 389                                ldlm_ns_name(ns));
 390                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
 391                 }
 392         }
 393
 394         return count;
 395 }
 396 LUSTRE_RW_ATTR(lru_size);
 397
 398 static ssize_t lru_max_age_show(struct kobject *kobj, struct attribute *attr,
 399                                 char *buf)
 400 {
 401         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 402                                                  ns_kobj);
 403
 404         return sprintf(buf, "%lld\n", ktime_to_ms(ns->ns_max_age));
 405 }
 406
 407 static ssize_t lru_max_age_store(struct kobject *kobj, struct attribute *attr,
 408                                  const char *buffer, size_t count)
 409 {
 410         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 411                                                  ns_kobj);
 412         int scale = NSEC_PER_MSEC;
 413         unsigned long long tmp;
 414         char *buf;
 415
 416         /* Did the user ask in seconds or milliseconds. Default is in ms */
 417         buf = strstr(buffer, "ms");
 418         if (!buf) {
 419                 buf = strchr(buffer, 's');
 420                 if (buf)
 421                         scale = NSEC_PER_SEC;
 422         }
 423
 424         if (buf)
 425                 *buf = '\0';
 426
 427         if (kstrtoull(buffer, 10, &tmp))
 428                 return -EINVAL;
 429
 430         ns->ns_max_age = ktime_set(0, tmp * scale);
 431
 432         return count;
 433 }
 434 LUSTRE_RW_ATTR(lru_max_age);
 435
 436 static ssize_t early_lock_cancel_show(struct kobject *kobj,
 437                                       struct attribute *attr,
 438                                       char *buf)
 439 {
 440         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 441                                                  ns_kobj);
 442
 443         return sprintf(buf, "%d\n", ns_connect_cancelset(ns));
 444 }
 445
 446 static ssize_t early_lock_cancel_store(struct kobject *kobj,
 447                                        struct attribute *attr,
 448                                        const char *buffer,
 449                                        size_t count)
 450 {
 451         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 452                                                  ns_kobj);
 453         unsigned long supp = -1;
 454         int rc;
 455
 456         rc = kstrtoul(buffer, 10, &supp);
 457         if (rc < 0)
 458                 return rc;
 459
 460         if (supp == 0)
 461                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
 462         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
 463                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
 464         return count;
 465 }
 466 LUSTRE_RW_ATTR(early_lock_cancel);
 467
 468 static ssize_t dirty_age_limit_show(struct kobject *kobj,
 469                                     struct attribute *attr, char *buf)
 470 {
 471         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 472                                                  ns_kobj);
 473
 474         return sprintf(buf, "%llu\n", ns->ns_dirty_age_limit);
 475 }
 476
 477 static ssize_t dirty_age_limit_store(struct kobject *kobj,
 478                                      struct attribute *attr,
 479                                      const char *buffer, size_t count)
 480 {
 481         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 482                                                  ns_kobj);
 483         unsigned long long tmp;
 484
 485         if (kstrtoull(buffer, 10, &tmp))
 486                 return -EINVAL;
 487
 488         ns->ns_dirty_age_limit = tmp;
 489
 490         return count;
 491 }
 492 LUSTRE_RW_ATTR(dirty_age_limit);
 493
 494 #ifdef HAVE_SERVER_SUPPORT
 495 static ssize_t ctime_age_limit_show(struct kobject *kobj,
 496                                     struct attribute *attr, char *buf)
 497 {
 498         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 499                                                  ns_kobj);
 500
 501         return sprintf(buf, "%llu\n", ns->ns_ctime_age_limit);
 502 }
 503
 504 static ssize_t ctime_age_limit_store(struct kobject *kobj,
 505                                      struct attribute *attr,
 506                                      const char *buffer, size_t count)
 507 {
 508         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 509                                                  ns_kobj);
 510         unsigned long long tmp;
 511
 512         if (kstrtoull(buffer, 10, &tmp))
 513                 return -EINVAL;
 514
 515         ns->ns_ctime_age_limit = tmp;
 516
 517         return count;
 518 }
 519 LUSTRE_RW_ATTR(ctime_age_limit);
 520
 521 static ssize_t lock_timeouts_show(struct kobject *kobj, struct attribute *attr,
 522                                   char *buf)
 523 {
 524         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 525                                                  ns_kobj);
 526
 527         return sprintf(buf, "%d\n", ns->ns_timeouts);
 528 }
 529 LUSTRE_RO_ATTR(lock_timeouts);
 530
 531 static ssize_t max_nolock_bytes_show(struct kobject *kobj,
 532                                      struct attribute *attr, char *buf)
 533 {
 534         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 535                                                  ns_kobj);
 536
 537         return sprintf(buf, "%u\n", ns->ns_max_nolock_size);
 538 }
 539
 540 static ssize_t max_nolock_bytes_store(struct kobject *kobj,
 541                                       struct attribute *attr,
 542                                       const char *buffer, size_t count)
 543 {
 544         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 545                                                  ns_kobj);
 546         unsigned long tmp;
 547         int err;
 548
 549         err = kstrtoul(buffer, 10, &tmp);
 550         if (err != 0)
 551                 return -EINVAL;
 552
 553         ns->ns_max_nolock_size = tmp;
 554
 555         return count;
 556 }
 557 LUSTRE_RW_ATTR(max_nolock_bytes);
 558
 559 static ssize_t contention_seconds_show(struct kobject *kobj,
 560                                        struct attribute *attr, char *buf)
 561 {
 562         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 563                                                  ns_kobj);
 564
 565         return sprintf(buf, "%llu\n", ns->ns_contention_time);
 566 }
 567
 568 static ssize_t contention_seconds_store(struct kobject *kobj,
 569                                         struct attribute *attr,
 570                                         const char *buffer, size_t count)
 571 {
 572         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 573                                                  ns_kobj);
 574         unsigned long long tmp;
 575
 576         if (kstrtoull(buffer, 10, &tmp))
 577                 return -EINVAL;
 578
 579         ns->ns_contention_time = tmp;
 580
 581         return count;
 582 }
 583 LUSTRE_RW_ATTR(contention_seconds);
 584
 585 static ssize_t contended_locks_show(struct kobject *kobj,
 586                                     struct attribute *attr, char *buf)
 587 {
 588         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 589                                                  ns_kobj);
 590
 591         return sprintf(buf, "%u\n", ns->ns_contended_locks);
 592 }
 593
 594 static ssize_t contended_locks_store(struct kobject *kobj,
 595                                      struct attribute *attr,
 596                                      const char *buffer, size_t count)
 597 {
 598         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 599                                                  ns_kobj);
 600         unsigned long tmp;
 601         int err;
 602
 603         err = kstrtoul(buffer, 10, &tmp);
 604         if (err != 0)
 605                 return -EINVAL;
 606
 607         ns->ns_contended_locks = tmp;
 608
 609         return count;
 610 }
 611 LUSTRE_RW_ATTR(contended_locks);
 612
 613 static ssize_t max_parallel_ast_show(struct kobject *kobj,
 614                                      struct attribute *attr, char *buf)
 615 {
 616         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 617                                                  ns_kobj);
 618
 619         return sprintf(buf, "%u\n", ns->ns_max_parallel_ast);
 620 }
 621
 622 static ssize_t max_parallel_ast_store(struct kobject *kobj,
 623                                       struct attribute *attr,
 624                                       const char *buffer, size_t count)
 625 {
 626         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 627                                                  ns_kobj);
 628         unsigned long tmp;
 629         int err;
 630
 631         err = kstrtoul(buffer, 10, &tmp);
 632         if (err != 0)
 633                 return -EINVAL;
 634
 635         ns->ns_max_parallel_ast = tmp;
 636
 637         return count;
 638 }
 639 LUSTRE_RW_ATTR(max_parallel_ast);
 640
 641 #endif /* HAVE_SERVER_SUPPORT */
 642
 643 /* These are for namespaces in /sys/fs/lustre/ldlm/namespaces/ */
 644 static struct attribute *ldlm_ns_attrs[] = {
 645         &lustre_attr_resource_count.attr,
 646         &lustre_attr_lock_count.attr,
 647         &lustre_attr_lock_unused_count.attr,
 648         &lustre_attr_lru_size.attr,
 649         &lustre_attr_lru_max_age.attr,
 650         &lustre_attr_early_lock_cancel.attr,
 651         &lustre_attr_dirty_age_limit.attr,
 652 #ifdef HAVE_SERVER_SUPPORT
 653         &lustre_attr_ctime_age_limit.attr,
 654         &lustre_attr_lock_timeouts.attr,
 655         &lustre_attr_max_nolock_bytes.attr,
 656         &lustre_attr_contention_seconds.attr,
 657         &lustre_attr_contended_locks.attr,
 658         &lustre_attr_max_parallel_ast.attr,
 659 #endif
 660         NULL,
 661 };
 662
 663 static void ldlm_ns_release(struct kobject *kobj)
 664 {
 665         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
 666                                                  ns_kobj);
 667         complete(&ns->ns_kobj_unregister);
 668 }
 669
 670 static struct kobj_type ldlm_ns_ktype = {
 671         .default_attrs  = ldlm_ns_attrs,
 672         .sysfs_ops      = &lustre_sysfs_ops,
 673         .release        = ldlm_ns_release,
 674 };
 675
 676 static void ldlm_namespace_debugfs_unregister(struct ldlm_namespace *ns)
 677 {
 678         if (IS_ERR_OR_NULL(ns->ns_debugfs_entry))
 679                 CERROR("dlm namespace %s has no procfs dir?\n",
 680                        ldlm_ns_name(ns));
 681         else
 682                 debugfs_remove_recursive(ns->ns_debugfs_entry);
 683
 684         if (ns->ns_stats != NULL)
 685                 lprocfs_free_stats(&ns->ns_stats);
 686 }
 687
 688 void ldlm_namespace_sysfs_unregister(struct ldlm_namespace *ns)
 689 {
 690         kobject_put(&ns->ns_kobj);
 691         wait_for_completion(&ns->ns_kobj_unregister);
 692 }
 693
 694 int ldlm_namespace_sysfs_register(struct ldlm_namespace *ns)
 695 {
 696         int err;
 697
 698         ns->ns_kobj.kset = ldlm_ns_kset;
 699         init_completion(&ns->ns_kobj_unregister);
 700         err = kobject_init_and_add(&ns->ns_kobj, &ldlm_ns_ktype, NULL,
 701                                    "%s", ldlm_ns_name(ns));
 702
 703         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
 704         if (!ns->ns_stats) {
 705                 kobject_put(&ns->ns_kobj);
 706                 return -ENOMEM;
 707         }
 708
 709         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
 710                              LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
 711
 712         return err;
 713 }
 714
 715 static int ldlm_namespace_debugfs_register(struct ldlm_namespace *ns)
 716 {
 717         struct dentry *ns_entry;
 718
 719         if (!IS_ERR_OR_NULL(ns->ns_debugfs_entry)) {
 720                 ns_entry = ns->ns_debugfs_entry;
 721         } else {
 722                 ns_entry = debugfs_create_dir(ldlm_ns_name(ns),
 723                                               ldlm_ns_debugfs_dir);
 724                 if (!ns_entry)
 725                         return -ENOMEM;
 726                 ns->ns_debugfs_entry = ns_entry;
 727         }
 728
 729         return 0;
 730 }
 731 #undef MAX_STRING_SIZE
 732
 733 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
 734                                   const void *key, unsigned int mask)
 735 {
 736         const struct ldlm_res_id *id = key;
 737         unsigned int val = 0;
 738         unsigned int i;
 739
 740         for (i = 0; i < RES_NAME_SIZE; i++)
 741                 val += id->name[i];
 742         return val & mask;
 743 }
 744
 745 static unsigned int ldlm_res_hop_fid_hash(const struct ldlm_res_id *id, unsigned int bits)
 746 {
 747         struct lu_fid       fid;
 748         __u32               hash;
 749         __u32               val;
 750
 751         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
 752         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
 753         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
 754
 755         hash = fid_flatten32(&fid);
 756         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 757         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
 758                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
 759         } else {
 760                 val = fid_oid(&fid);
 761         }
 762         hash += (val >> 5) + (val << 11);
 763         return cfs_hash_32(hash, bits);
 764 }
 765
 766 static void *ldlm_res_hop_key(struct hlist_node *hnode)
 767 {
 768         struct ldlm_resource   *res;
 769
 770         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 771         return &res->lr_name;
 772 }
 773
 774 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
 775 {
 776         struct ldlm_resource   *res;
 777
 778         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 779         return ldlm_res_eq((const struct ldlm_res_id *)key,
 780                            (const struct ldlm_res_id *)&res->lr_name);
 781 }
 782
 783 static void *ldlm_res_hop_object(struct hlist_node *hnode)
 784 {
 785         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
 786 }
 787
 788 static void
 789 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 790 {
 791         struct ldlm_resource *res;
 792
 793         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 794         ldlm_resource_getref(res);
 795 }
 796
 797 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 798 {
 799         struct ldlm_resource *res;
 800
 801         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
 802         ldlm_resource_putref(res);
 803 }
 804
 805 static struct cfs_hash_ops ldlm_ns_hash_ops = {
 806         .hs_hash        = ldlm_res_hop_hash,
 807         .hs_key         = ldlm_res_hop_key,
 808         .hs_keycmp      = ldlm_res_hop_keycmp,
 809         .hs_keycpy      = NULL,
 810         .hs_object      = ldlm_res_hop_object,
 811         .hs_get         = ldlm_res_hop_get_locked,
 812         .hs_put         = ldlm_res_hop_put
 813 };
 814
 815 static struct {
 816         /** hash bucket bits */
 817         unsigned                nsd_bkt_bits;
 818         /** hash bits */
 819         unsigned                nsd_all_bits;
 820 } ldlm_ns_hash_defs[] = {
 821         [LDLM_NS_TYPE_MDC] = {
 822                 .nsd_bkt_bits   = 11,
 823                 .nsd_all_bits   = 16,
 824         },
 825         [LDLM_NS_TYPE_MDT] = {
 826                 .nsd_bkt_bits   = 14,
 827                 .nsd_all_bits   = 21,
 828         },
 829         [LDLM_NS_TYPE_OSC] = {
 830                 .nsd_bkt_bits   = 8,
 831                 .nsd_all_bits   = 12,
 832         },
 833         [LDLM_NS_TYPE_OST] = {
 834                 .nsd_bkt_bits   = 11,
 835                 .nsd_all_bits   = 17,
 836         },
 837         [LDLM_NS_TYPE_MGC] = {
 838                 .nsd_bkt_bits   = 3,
 839                 .nsd_all_bits   = 4,
 840         },
 841         [LDLM_NS_TYPE_MGT] = {
 842                 .nsd_bkt_bits   = 3,
 843                 .nsd_all_bits   = 4,
 844         },
 845 };
 846
 847 /**
 848  * Create and initialize new empty namespace.
 849  */
 850 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
 851                                           enum ldlm_side client,
 852                                           enum ldlm_appetite apt,
 853                                           enum ldlm_ns_type ns_type)
 854 {
 855         struct ldlm_namespace *ns = NULL;
 856         int idx;
 857         int rc;
 858
 859         ENTRY;
 860         LASSERT(obd != NULL);
 861
 862         rc = ldlm_get_ref();
 863         if (rc) {
 864                 CERROR("ldlm_get_ref failed: %d\n", rc);
 865                 RETURN(NULL);
 866         }
 867
 868         if (ns_type >= ARRAY_SIZE(ldlm_ns_hash_defs) ||
 869             ldlm_ns_hash_defs[ns_type].nsd_bkt_bits == 0) {
 870                 CERROR("Unknown type %d for ns %s\n", ns_type, name);
 871                 GOTO(out_ref, NULL);
 872         }
 873
 874         OBD_ALLOC_PTR(ns);
 875         if (!ns)
 876                 GOTO(out_ref, NULL);
 877
 878         ns->ns_rs_hash = cfs_hash_create(name,
 879                                          ldlm_ns_hash_defs[ns_type].nsd_all_bits,
 880                                          ldlm_ns_hash_defs[ns_type].nsd_all_bits,
 881                                          ldlm_ns_hash_defs[ns_type].nsd_bkt_bits,
 882                                          0,
 883                                          CFS_HASH_MIN_THETA,
 884                                          CFS_HASH_MAX_THETA,
 885                                          &ldlm_ns_hash_ops,
 886                                          CFS_HASH_DEPTH |
 887                                          CFS_HASH_BIGNAME |
 888                                          CFS_HASH_SPIN_BKTLOCK |
 889                                          CFS_HASH_NO_ITEMREF);
 890         if (ns->ns_rs_hash == NULL)
 891                 GOTO(out_ns, NULL);
 892
 893         ns->ns_bucket_bits = ldlm_ns_hash_defs[ns_type].nsd_all_bits -
 894                              ldlm_ns_hash_defs[ns_type].nsd_bkt_bits;
 895
 896         OBD_ALLOC_LARGE(ns->ns_rs_buckets,
 897                         BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0]));
 898         if (!ns->ns_rs_buckets)
 899                 goto out_hash;
 900
 901         for (idx = 0; idx < (1 << ns->ns_bucket_bits); idx++) {
 902                 struct ldlm_ns_bucket *nsb = &ns->ns_rs_buckets[idx];
 903
 904                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
 905                 nsb->nsb_namespace = ns;
 906                 nsb->nsb_reclaim_start = 0;
 907                 atomic_set(&nsb->nsb_count, 0);
 908         }
 909
 910         ns->ns_obd = obd;
 911         ns->ns_appetite = apt;
 912         ns->ns_client = client;
 913         ns->ns_name = kstrdup(name, GFP_KERNEL);
 914         if (!ns->ns_name)
 915                 goto out_hash;
 916
 917         INIT_LIST_HEAD(&ns->ns_list_chain);
 918         INIT_LIST_HEAD(&ns->ns_unused_list);
 919         spin_lock_init(&ns->ns_lock);
 920         atomic_set(&ns->ns_bref, 0);
 921         init_waitqueue_head(&ns->ns_waitq);
 922
 923         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
 924         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
 925         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
 926
 927         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
 928         ns->ns_nr_unused          = 0;
 929         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
 930         ns->ns_max_age            = ktime_set(LDLM_DEFAULT_MAX_ALIVE, 0);
 931         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
 932         ns->ns_dirty_age_limit    = LDLM_DIRTY_AGE_LIMIT;
 933         ns->ns_timeouts           = 0;
 934         ns->ns_orig_connect_flags = 0;
 935         ns->ns_connect_flags      = 0;
 936         ns->ns_stopping           = 0;
 937         ns->ns_reclaim_start      = 0;
 938         ns->ns_last_pos           = &ns->ns_unused_list;
 939
 940         rc = ldlm_namespace_sysfs_register(ns);
 941         if (rc) {
 942                 CERROR("Can't initialize ns sysfs, rc %d\n", rc);
 943                 GOTO(out_hash, rc);
 944         }
 945
 946         rc = ldlm_namespace_debugfs_register(ns);
 947         if (rc) {
 948                 CERROR("Can't initialize ns proc, rc %d\n", rc);
 949                 GOTO(out_sysfs, rc);
 950         }
 951
 952         idx = ldlm_namespace_nr_read(client);
 953         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
 954         if (rc) {
 955                 CERROR("Can't initialize lock pool, rc %d\n", rc);
 956                 GOTO(out_proc, rc);
 957         }
 958
 959         ldlm_namespace_register(ns, client);
 960         RETURN(ns);
 961 out_proc:
 962         ldlm_namespace_debugfs_unregister(ns);
 963 out_sysfs:
 964         ldlm_namespace_sysfs_unregister(ns);
 965         ldlm_namespace_cleanup(ns, 0);
 966 out_hash:
 967         OBD_FREE_LARGE(ns->ns_rs_buckets,
 968                        BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0]));
 969         kfree(ns->ns_name);
 970         cfs_hash_putref(ns->ns_rs_hash);
 971 out_ns:
 972         OBD_FREE_PTR(ns);
 973 out_ref:
 974         ldlm_put_ref();
 975         RETURN(NULL);
 976 }
 977 EXPORT_SYMBOL(ldlm_namespace_new);
 978
 979 /**
 980  * Cancel and destroy all locks on a resource.
 981  *
 982  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
 983  * clean up.  This is currently only used for recovery, and we make
 984  * certain assumptions as a result--notably, that we shouldn't cancel
 985  * locks with refs.
 986  */
 987 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 988                              __u64 flags)
 989 {
 990         struct list_head *tmp;
 991         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
 992         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
 993
 994         do {
 995                 struct ldlm_lock *lock = NULL;
 996
 997                 /* First, we look for non-cleaned-yet lock
 998                  * all cleaned locks are marked by CLEANED flag. */
 999                 lock_res(res);
1000                 list_for_each(tmp, q) {
1001                         lock = list_entry(tmp, struct ldlm_lock,
1002                                           l_res_link);
1003                         if (ldlm_is_cleaned(lock)) {
1004                                 lock = NULL;
1005                                 continue;
1006                         }
1007                         LDLM_LOCK_GET(lock);
1008                         ldlm_set_cleaned(lock);
1009                         break;
1010                 }
1011
1012                 if (lock == NULL) {
1013                         unlock_res(res);
1014                         break;
1015                 }
1016
1017                 /* Set CBPENDING so nothing in the cancellation path
1018                  * can match this lock. */
1019                 ldlm_set_cbpending(lock);
1020                 ldlm_set_failed(lock);
1021                 lock->l_flags |= flags;
1022
1023                 /* ... without sending a CANCEL message for local_only. */
1024                 if (local_only)
1025                         ldlm_set_local_only(lock);
1026
1027                 if (local_only && (lock->l_readers || lock->l_writers)) {
1028                         /*
1029                          * This is a little bit gross, but much better than the
1030                          * alternative: pretend that we got a blocking AST from
1031                          * the server, so that when the lock is decref'd, it
1032                          * will go away ...
1033                          */
1034                         unlock_res(res);
1035                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
1036                         if (lock->l_flags & LDLM_FL_FAIL_LOC)
1037                                 schedule_timeout_uninterruptible(
1038                                         cfs_time_seconds(4));
1039
1040                         if (lock->l_completion_ast)
1041                                 lock->l_completion_ast(lock,
1042                                                        LDLM_FL_FAILED, NULL);
1043                         LDLM_LOCK_RELEASE(lock);
1044                         continue;
1045                 }
1046
1047                 if (client) {
1048                         struct lustre_handle lockh;
1049
1050                         unlock_res(res);
1051                         ldlm_lock2handle(lock, &lockh);
1052                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
1053                         if (rc)
1054                                 CERROR("ldlm_cli_cancel: %d\n", rc);
1055                 } else {
1056                         unlock_res(res);
1057                         LDLM_DEBUG(lock,
1058                                    "Freeing a lock still held by a client node");
1059                         ldlm_lock_cancel(lock);
1060                 }
1061                 LDLM_LOCK_RELEASE(lock);
1062         } while (1);
1063 }
1064
1065 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1066                                struct hlist_node *hnode, void *arg)
1067 {
1068         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1069         __u64 flags = *(__u64 *)arg;
1070
1071         cleanup_resource(res, &res->lr_granted, flags);
1072         cleanup_resource(res, &res->lr_waiting, flags);
1073
1074         return 0;
1075 }
1076
1077 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1078                                   struct hlist_node *hnode, void *arg)
1079 {
1080         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
1081
1082         lock_res(res);
1083         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
1084                "(%d) after lock cleanup; forcing cleanup.\n",
1085                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
1086                atomic_read(&res->lr_refcount) - 1);
1087
1088         /* Use D_NETERROR since it is in the default mask */
1089         ldlm_resource_dump(D_NETERROR, res);
1090         unlock_res(res);
1091         return 0;
1092 }
1093
1094 /**
1095  * Cancel and destroy all locks in the namespace.
1096  *
1097  * Typically used during evictions when server notified client that it was
1098  * evicted and all of its state needs to be destroyed.
1099  * Also used during shutdown.
1100  */
1101 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
1102 {
1103         if (ns == NULL) {
1104                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
1105                 return ELDLM_OK;
1106         }
1107
1108         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
1109                                  &flags, 0);
1110         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
1111                                  NULL, 0);
1112         return ELDLM_OK;
1113 }
1114 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1115
1116 /**
1117  * Attempts to free namespace.
1118  *
1119  * Only used when namespace goes away, like during an unmount.
1120  */
1121 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
1122 {
1123         ENTRY;
1124
1125         /* At shutdown time, don't call the cancellation callback */
1126         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
1127
1128         if (atomic_read(&ns->ns_bref) > 0) {
1129                 int rc;
1130                 CDEBUG(D_DLMTRACE,
1131                        "dlm namespace %s free waiting on refcount %d\n",
1132                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
1133 force_wait:
1134                 if (force)
1135                         rc = wait_event_idle_timeout(
1136                                 ns->ns_waitq,
1137                                 atomic_read(&ns->ns_bref) == 0,
1138                                 cfs_time_seconds(1) / 4);
1139                 else
1140                         rc = l_wait_event_abortable(
1141                                 ns->ns_waitq, atomic_read(&ns->ns_bref) == 0);
1142
1143                 /* Forced cleanups should be able to reclaim all references,
1144                  * so it's safe to wait forever... we can't leak locks... */
1145                 if (force && rc == 0) {
1146                         rc = -ETIMEDOUT;
1147                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
1148                                        "namespace with %d resources in use, "
1149                                        "(rc=%d)\n", ldlm_ns_name(ns),
1150                                        atomic_read(&ns->ns_bref), rc);
1151                         GOTO(force_wait, rc);
1152                 }
1153
1154                 if (atomic_read(&ns->ns_bref)) {
1155                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
1156                                        "with %d resources in use, (rc=%d)\n",
1157                                        ldlm_ns_name(ns),
1158                                        atomic_read(&ns->ns_bref), rc);
1159                         RETURN(ELDLM_NAMESPACE_EXISTS);
1160                 }
1161                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
1162                        ldlm_ns_name(ns));
1163         }
1164
1165         RETURN(ELDLM_OK);
1166 }
1167
1168 /**
1169  * Performs various cleanups for passed \a ns to make it drop refc and be
1170  * ready for freeing. Waits for refc == 0.
1171  *
1172  * The following is done:
1173  * (0) Unregister \a ns from its list to make inaccessible for potential
1174  * users like pools thread and others;
1175  * (1) Clear all locks in \a ns.
1176  */
1177 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
1178                                struct obd_import *imp,
1179                                int force)
1180 {
1181         int rc;
1182
1183         ENTRY;
1184         if (!ns) {
1185                 EXIT;
1186                 return;
1187         }
1188
1189         spin_lock(&ns->ns_lock);
1190         ns->ns_stopping = 1;
1191         spin_unlock(&ns->ns_lock);
1192
1193         /*
1194          * Can fail with -EINTR when force == 0 in which case try harder.
1195          */
1196         rc = __ldlm_namespace_free(ns, force);
1197         if (rc != ELDLM_OK) {
1198                 if (imp) {
1199                         ptlrpc_disconnect_import(imp, 0);
1200                         ptlrpc_invalidate_import(imp);
1201                 }
1202
1203                 /*
1204                  * With all requests dropped and the import inactive
1205                  * we are gaurenteed all reference will be dropped.
1206                  */
1207                 rc = __ldlm_namespace_free(ns, 1);
1208                 LASSERT(rc == 0);
1209         }
1210         EXIT;
1211 }
1212 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1213
1214 /**
1215  * Performs freeing memory structures related to \a ns. This is only done
1216  * when ldlm_namespce_free_prior() successfully removed all resources
1217  * referencing \a ns and its refc == 0.
1218  */
1219 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1220 {
1221         ENTRY;
1222         if (!ns) {
1223                 EXIT;
1224                 return;
1225         }
1226
1227         /* Make sure that nobody can find this ns in its list. */
1228         ldlm_namespace_unregister(ns, ns->ns_client);
1229         /* Fini pool _before_ parent proc dir is removed. This is important as
1230          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1231          * Removing it after @dir may cause oops. */
1232         ldlm_pool_fini(&ns->ns_pool);
1233
1234         ldlm_namespace_debugfs_unregister(ns);
1235         ldlm_namespace_sysfs_unregister(ns);
1236         cfs_hash_putref(ns->ns_rs_hash);
1237         OBD_FREE_LARGE(ns->ns_rs_buckets,
1238                        BIT(ns->ns_bucket_bits) * sizeof(ns->ns_rs_buckets[0]));
1239         kfree(ns->ns_name);
1240         /* Namespace \a ns should be not on list at this time, otherwise
1241          * this will cause issues related to using freed \a ns in poold
1242          * thread.
1243          */
1244         LASSERT(list_empty(&ns->ns_list_chain));
1245         OBD_FREE_PTR(ns);
1246         ldlm_put_ref();
1247         EXIT;
1248 }
1249 EXPORT_SYMBOL(ldlm_namespace_free_post);
1250
1251 /**
1252  * Cleanup the resource, and free namespace.
1253  * bug 12864:
1254  * Deadlock issue:
1255  * proc1: destroy import
1256  *        class_disconnect_export(grab cl_sem) ->
1257  *              -> ldlm_namespace_free ->
1258  *              -> lprocfs_remove(grab _lprocfs_lock).
1259  * proc2: read proc info
1260  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1261  *              -> osc_rd_active, etc(grab cl_sem).
1262  *
1263  * So that I have to split the ldlm_namespace_free into two parts - the first
1264  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1265  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1266  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1267  * held.
1268  */
1269 void ldlm_namespace_free(struct ldlm_namespace *ns,
1270                          struct obd_import *imp,
1271                          int force)
1272 {
1273         ldlm_namespace_free_prior(ns, imp, force);
1274         ldlm_namespace_free_post(ns);
1275 }
1276 EXPORT_SYMBOL(ldlm_namespace_free);
1277
1278 void ldlm_namespace_get(struct ldlm_namespace *ns)
1279 {
1280         atomic_inc(&ns->ns_bref);
1281 }
1282
1283 /* This is only for callers that care about refcount */
1284 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1285 {
1286         return atomic_inc_return(&ns->ns_bref);
1287 }
1288
1289 void ldlm_namespace_put(struct ldlm_namespace *ns)
1290 {
1291         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1292                 wake_up(&ns->ns_waitq);
1293                 spin_unlock(&ns->ns_lock);
1294         }
1295 }
1296
1297 /** Register \a ns in the list of namespaces */
1298 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1299 {
1300         mutex_lock(ldlm_namespace_lock(client));
1301         LASSERT(list_empty(&ns->ns_list_chain));
1302         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1303         ldlm_namespace_nr_inc(client);
1304         mutex_unlock(ldlm_namespace_lock(client));
1305 }
1306
1307 /** Unregister \a ns from the list of namespaces. */
1308 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1309 {
1310         mutex_lock(ldlm_namespace_lock(client));
1311         LASSERT(!list_empty(&ns->ns_list_chain));
1312         /* Some asserts and possibly other parts of the code are still
1313          * using list_empty(&ns->ns_list_chain). This is why it is
1314          * important to use list_del_init() here. */
1315         list_del_init(&ns->ns_list_chain);
1316         ldlm_namespace_nr_dec(client);
1317         mutex_unlock(ldlm_namespace_lock(client));
1318 }
1319
1320 /** Should be called with ldlm_namespace_lock(client) taken. */
1321 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1322                                           enum ldlm_side client)
1323 {
1324         LASSERT(!list_empty(&ns->ns_list_chain));
1325         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1326         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1327 }
1328
1329 /** Should be called with ldlm_namespace_lock(client) taken. */
1330 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1331                                             enum ldlm_side client)
1332 {
1333         LASSERT(!list_empty(&ns->ns_list_chain));
1334         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1335         list_move_tail(&ns->ns_list_chain,
1336                        ldlm_namespace_inactive_list(client));
1337 }
1338
1339 /** Should be called with ldlm_namespace_lock(client) taken. */
1340 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1341 {
1342         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1343         LASSERT(!list_empty(ldlm_namespace_list(client)));
1344         return container_of(ldlm_namespace_list(client)->next,
1345                             struct ldlm_namespace, ns_list_chain);
1346 }
1347
1348 static bool ldlm_resource_extent_new(struct ldlm_resource *res)
1349 {
1350         int idx;
1351
1352         OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1353                        sizeof(*res->lr_itree) * LCK_MODE_NUM);
1354         if (res->lr_itree == NULL)
1355                 return false;
1356         /* Initialize interval trees for each lock mode. */
1357         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1358                 res->lr_itree[idx].lit_size = 0;
1359                 res->lr_itree[idx].lit_mode = 1 << idx;
1360                 res->lr_itree[idx].lit_root = NULL;
1361         }
1362         return true;
1363 }
1364
1365 static bool ldlm_resource_inodebits_new(struct ldlm_resource *res)
1366 {
1367         int i;
1368
1369         OBD_ALLOC_PTR(res->lr_ibits_queues);
1370         if (res->lr_ibits_queues == NULL)
1371                 return false;
1372         for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
1373                 INIT_LIST_HEAD(&res->lr_ibits_queues->liq_waiting[i]);
1374         return true;
1375 }
1376
1377 /** Create and initialize new resource. */
1378 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1379 {
1380         struct ldlm_resource *res;
1381         bool rc;
1382
1383         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1384         if (res == NULL)
1385                 return NULL;
1386
1387         switch (ldlm_type) {
1388         case LDLM_EXTENT:
1389                 rc = ldlm_resource_extent_new(res);
1390                 break;
1391         case LDLM_IBITS:
1392                 rc = ldlm_resource_inodebits_new(res);
1393                 break;
1394         default:
1395                 rc = true;
1396                 break;
1397         }
1398         if (!rc) {
1399                 OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1400                 return NULL;
1401         }
1402
1403         INIT_LIST_HEAD(&res->lr_granted);
1404         INIT_LIST_HEAD(&res->lr_waiting);
1405
1406         atomic_set(&res->lr_refcount, 1);
1407         spin_lock_init(&res->lr_lock);
1408         lu_ref_init(&res->lr_reference);
1409
1410         /* Since LVB init can be delayed now, there is no longer need to
1411          * immediatelly acquire mutex here. */
1412         mutex_init(&res->lr_lvb_mutex);
1413         res->lr_lvb_initialized = false;
1414
1415         return res;
1416 }
1417
1418 static void ldlm_resource_free(struct ldlm_resource *res)
1419 {
1420         if (res->lr_type == LDLM_EXTENT) {
1421                 if (res->lr_itree != NULL)
1422                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1423                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1424         } else if (res->lr_type == LDLM_IBITS) {
1425                 if (res->lr_ibits_queues != NULL)
1426                         OBD_FREE_PTR(res->lr_ibits_queues);
1427         }
1428
1429         OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1430 }
1431
1432 /**
1433  * Return a reference to resource with given name, creating it if necessary.
1434  * Args: namespace with ns_lock unlocked
1435  * Locks: takes and releases NS hash-lock and res->lr_lock
1436  * Returns: referenced, unlocked ldlm_resource or NULL
1437  */
1438 struct ldlm_resource *
1439 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1440                   const struct ldlm_res_id *name, enum ldlm_type type,
1441                   int create)
1442 {
1443         struct hlist_node       *hnode;
1444         struct ldlm_resource    *res = NULL;
1445         struct cfs_hash_bd              bd;
1446         __u64                   version;
1447         int                     ns_refcount = 0;
1448         int hash;
1449
1450         LASSERT(ns != NULL);
1451         LASSERT(parent == NULL);
1452         LASSERT(ns->ns_rs_hash != NULL);
1453         LASSERT(name->name[0] != 0);
1454
1455         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1456         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1457         if (hnode != NULL) {
1458                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1459                 GOTO(found, res);
1460         }
1461
1462         version = cfs_hash_bd_version_get(&bd);
1463         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1464
1465         if (create == 0)
1466                 return ERR_PTR(-ENOENT);
1467
1468         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1469                  "type: %d\n", type);
1470         res = ldlm_resource_new(type);
1471         if (res == NULL)
1472                 return ERR_PTR(-ENOMEM);
1473
1474         hash = ldlm_res_hop_fid_hash(name, ns->ns_bucket_bits);
1475         res->lr_ns_bucket = &ns->ns_rs_buckets[hash];
1476         res->lr_name = *name;
1477         res->lr_type = type;
1478
1479         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1480         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1481                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1482
1483         if (hnode != NULL) {
1484                 /* Someone won the race and already added the resource. */
1485                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1486                 /* Clean lu_ref for failed resource. */
1487                 lu_ref_fini(&res->lr_reference);
1488                 ldlm_resource_free(res);
1489 found:
1490                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1491                 return res;
1492         }
1493         /* We won! Let's add the resource. */
1494         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1495         if (atomic_inc_return(&res->lr_ns_bucket->nsb_count) == 1)
1496                 ns_refcount = ldlm_namespace_get_return(ns);
1497
1498         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1499
1500         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1501
1502         /* Let's see if we happened to be the very first resource in this
1503          * namespace. If so, and this is a client namespace, we need to move
1504          * the namespace into the active namespaces list to be patrolled by
1505          * the ldlm_poold. */
1506         if (ns_is_client(ns) && ns_refcount == 1) {
1507                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1508                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1509                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1510         }
1511
1512         return res;
1513 }
1514 EXPORT_SYMBOL(ldlm_resource_get);
1515
1516 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1517 {
1518         LASSERT(res != NULL);
1519         LASSERT(res != LP_POISON);
1520         atomic_inc(&res->lr_refcount);
1521         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1522                atomic_read(&res->lr_refcount));
1523         return res;
1524 }
1525
1526 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1527                                          struct ldlm_resource *res)
1528 {
1529         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1530
1531         if (!list_empty(&res->lr_granted)) {
1532                 ldlm_resource_dump(D_ERROR, res);
1533                 LBUG();
1534         }
1535
1536         if (!list_empty(&res->lr_waiting)) {
1537                 ldlm_resource_dump(D_ERROR, res);
1538                 LBUG();
1539         }
1540
1541         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1542                                bd, &res->lr_hash);
1543         lu_ref_fini(&res->lr_reference);
1544         if (atomic_dec_and_test(&nsb->nsb_count))
1545                 ldlm_namespace_put(nsb->nsb_namespace);
1546 }
1547
1548 /* Returns 1 if the resource was freed, 0 if it remains. */
1549 int ldlm_resource_putref(struct ldlm_resource *res)
1550 {
1551         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1552         struct cfs_hash_bd   bd;
1553
1554         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1555         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1556                res, atomic_read(&res->lr_refcount) - 1);
1557
1558         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1559         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1560                 __ldlm_resource_putref_final(&bd, res);
1561                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1562                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1563                         ns->ns_lvbo->lvbo_free(res);
1564                 ldlm_resource_free(res);
1565                 return 1;
1566         }
1567         return 0;
1568 }
1569 EXPORT_SYMBOL(ldlm_resource_putref);
1570
1571 /**
1572  * Add a lock into a given resource into specified lock list.
1573  */
1574 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1575                             struct ldlm_lock *lock)
1576 {
1577         check_res_locked(res);
1578
1579         LDLM_DEBUG(lock, "About to add this lock");
1580
1581         if (ldlm_is_destroyed(lock)) {
1582                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1583                 return;
1584         }
1585
1586         LASSERT(list_empty(&lock->l_res_link));
1587
1588         list_add_tail(&lock->l_res_link, head);
1589
1590         if (res->lr_type == LDLM_IBITS)
1591                 ldlm_inodebits_add_lock(res, head, lock);
1592 }
1593
1594 /**
1595  * Insert a lock into resource after specified lock.
1596  *
1597  * Obtain resource description from the lock we are inserting after.
1598  */
1599 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1600                                      struct ldlm_lock *new)
1601 {
1602         struct ldlm_resource *res = original->l_resource;
1603
1604         check_res_locked(res);
1605
1606         ldlm_resource_dump(D_INFO, res);
1607         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1608
1609         if (ldlm_is_destroyed(new)) {
1610                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1611                 goto out;
1612         }
1613
1614         LASSERT(list_empty(&new->l_res_link));
1615
1616         list_add(&new->l_res_link, &original->l_res_link);
1617  out:;
1618 }
1619
1620 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1621 {
1622         int type = lock->l_resource->lr_type;
1623
1624         check_res_locked(lock->l_resource);
1625         switch (type) {
1626         case LDLM_PLAIN:
1627                 ldlm_unlink_lock_skiplist(lock);
1628                 break;
1629         case LDLM_EXTENT:
1630                 ldlm_extent_unlink_lock(lock);
1631                 break;
1632         case LDLM_IBITS:
1633                 ldlm_inodebits_unlink_lock(lock);
1634                 break;
1635         }
1636         list_del_init(&lock->l_res_link);
1637 }
1638 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1639
1640 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1641 {
1642         desc->lr_type = res->lr_type;
1643         desc->lr_name = res->lr_name;
1644 }
1645
1646 /**
1647  * Print information about all locks in all namespaces on this node to debug
1648  * log.
1649  */
1650 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1651 {
1652         struct list_head *tmp;
1653
1654         if (!((libcfs_debug | D_ERROR) & level))
1655                 return;
1656
1657         mutex_lock(ldlm_namespace_lock(client));
1658
1659         list_for_each(tmp, ldlm_namespace_list(client)) {
1660                 struct ldlm_namespace *ns;
1661
1662                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1663                 ldlm_namespace_dump(level, ns);
1664         }
1665
1666         mutex_unlock(ldlm_namespace_lock(client));
1667 }
1668
1669 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1670                               struct hlist_node *hnode, void *arg)
1671 {
1672         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1673         int    level = (int)(unsigned long)arg;
1674
1675         lock_res(res);
1676         ldlm_resource_dump(level, res);
1677         unlock_res(res);
1678
1679         return 0;
1680 }
1681
1682 /**
1683  * Print information about all locks in this namespace on this node to debug
1684  * log.
1685  */
1686 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1687 {
1688         if (!((libcfs_debug | D_ERROR) & level))
1689                 return;
1690
1691         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1692                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1693                ns_is_client(ns) ? "client" : "server");
1694
1695         if (ktime_get_seconds() < ns->ns_next_dump)
1696                 return;
1697
1698         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1699                                  ldlm_res_hash_dump,
1700                                  (void *)(unsigned long)level, 0);
1701         spin_lock(&ns->ns_lock);
1702         ns->ns_next_dump = ktime_get_seconds() + 10;
1703         spin_unlock(&ns->ns_lock);
1704 }
1705
1706 /**
1707  * Print information about all locks in this resource to debug log.
1708  */
1709 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1710 {
1711         struct ldlm_lock *lock;
1712         unsigned int granted = 0;
1713
1714         BUILD_BUG_ON(RES_NAME_SIZE != 4);
1715
1716         if (!((libcfs_debug | D_ERROR) & level))
1717                 return;
1718
1719         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1720                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1721
1722         if (!list_empty(&res->lr_granted)) {
1723                 CDEBUG(level, "Granted locks (in reverse order):\n");
1724                 list_for_each_entry_reverse(lock, &res->lr_granted,
1725                                                 l_res_link) {
1726                         LDLM_DEBUG_LIMIT(level, lock, "###");
1727                         if (!(level & D_CANTMASK) &&
1728                             ++granted > ldlm_dump_granted_max) {
1729                                 CDEBUG(level,
1730                                        "only dump %d granted locks to avoid DDOS.\n",
1731                                        granted);
1732                                 break;
1733                         }
1734                 }
1735         }
1736
1737         if (!list_empty(&res->lr_waiting)) {
1738                 CDEBUG(level, "Waiting locks:\n");
1739                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1740                         LDLM_DEBUG_LIMIT(level, lock, "###");
1741         }
1742 }
1743 EXPORT_SYMBOL(ldlm_resource_dump);